Queued job auto-cancelled ~1s after run start on self-hosted runner pool (no actor, no dispatch)

[jperdomo]

Queued job auto-cancelled ~1s after run start on self-hosted runner pool (no actor, no dispatch)

Summary

A job that is never dispatched to a runner is cancelled by GitHub about 1
second after the run is created. The cancelled job shows conclusion: cancelled, runner_id: 0, an empty steps array, canceled_by: null,
and conclusion_reason: null. It never touches a runner. The rest of the
run usually continues and every other job succeeds, but the run-level
conclusion ends cancelled, not failed.

Deterministic: 8 of 8 ci.yml runs on main since 2026-09-07
01:39 UTC show the kill, on push,and schedule triggers, across three
ci.yml revisions (including after we changed concurrency groups,and after
our two workarounds merged) Re-running does not help: two of the observed
runs are attempt-2 runs,and the same job cancelled again about 1-3s
after the re-run was created.

Reproduction (observed, 8/8)

Repo: jperdomo/gamma-origin, workflow ci.yml, job version-lockstep
(mostly( or detect-changes. The pool gamma-build is self-hosted, all
jobs runs-on: gamma-build.

run id	event	attempt	cancelled jobs	running jobs at death
34073748195	push	2	detect-changes (attempt 2, 02:15:24)	0
34127994494	schedule	1	version-lockstep (13:32:53) + 16 of 17 audit-deps legs (13:33:06)	0 at every death
34132649526	push	1	version-lockstep (14:23:05)	0
34157082412	push	1	version-lockstep (19:50:04)	0
34159056779	push	2	version-lockstep (20:42:30, attempt 2)	0
34164864812	push	1	version-lockstep (21:55:40)	0

In run 34164864812 a second job, warm-buildcache (app: also ended
cancelled (22:02:31) but it had been dispatched (runner_id=7,
steps=12( and died while finishing a different shape, post-dispatch,
not the queued kill this report covers.

What we checked,and ruled out

1. Same-run sibling concurrency race. We previously had all jobs share
   one concurrency group with cancel-in-progress: true. We moved to
   per-job groups (ci-${{ github.job }}-${{ github.ref }}, PR Github actions is not able to pull artifacts from github maven repo after runner upgrade from 2.305.0 to 2.306.0 #2692( and
   the auto-cancel continued unchanged, so the cancel is not caused by our
   concurrency configuration

2. Newer run superseding. At each cancel instant there is NO other
   ci.yml run in progress or queued. A whole-day audit shows exactly one
   ci.yml run per incident SHA

3. Human or bot cancel. canceled_by is null in the job payload; no
   user or app actor appears. Nothing else in the window calls the cancel API
   (our queued-run-sweep cron runs with actions: read and cannot cancel)

4. Re-run. Re-runs reproduce: attempt-2 runs cancelled the same
   job again about 1-3s after re-run creation. A human gh run cancel
   cannot explain that (it would need to know the attempt was re-created
   within seconds)**

5. Runner availability. Other jobs on the same pool dispatch,and run
   normally in the same run (build-images matrix, audit-deps,
   update-tags all succeed) At each queued-kill instant zero jobs were
   running against the pool`s 14 registered runners

6. GitHub-internal queued-cancel is the only model left. Discriminating
   cells:the victim is the first-queued job, invariant signature, recursion
   in re-runs,recursion across concurrency changes,and zero running jobs at
   every death. A known bug class matches: self-hosted jobs stuck or
   cancelled at runner_id=0 with an online,idle runner (actions/runner
   Self-Hosted Runner does not pick up queued jobs #3478, Self-hosted runner stuck on "Waiting for a runner to pick up this job..." in multi-step jobs #3609;community #186811, #186208).

Expected behavior

A queued job should stay queued until a runner is free, then dispatch,or,
if cancelled, the cancel should carry a reason/actor (conclusion_reason,
canceled_by) so automation can distinguish it from a deliberate cancel.

Actual behavior

The job is marked cancelled about 0-1s after run creation, before any
runner accepts it, with no reason,and no actor. There is currently no
reliable field on the job/run API to distinguish this internal cancel from a
human gh run cancel (our workaround compares started_at - run.created_at)

Impact

The cancelled job is a release-critical assertion; a downstream release gate
had to special-case the exact signature,and our CI now pages on it. Any repo
using a self-hosted pool for a small early job can hit this,and misread it as a
deliberate cancel

Environment

• Actions runner: self-hosted, Linux, pool gamma-build

Questions for the team

• Is there a known issue with queued-job cancellation on self-hosted pools
  during peak queue times?
• Is there a field (now or planned( that records an internal cancel reason, so
  automation can distinguish it from a user cancel without timing heuristics?
• Can we get the Actions queue/dispatch audit log for job 101873798915 (run
  34164864812, version-lockstep) from GitHub Support? A concurrency cancel
  would name the triggering run_id;a run cancel would show a run-level event;
  an internal bug shows no actor,and no reason,the expected outcome here.